Since the remarkable British Interplanetary Society starship study of the late 1970s - Daedalus - there have been significant developments in the areas of artificial intelligence and robotics. These will be critical technologies for any starship as indeed they are for the current generation of exploratory spacecraft and in-situ planetary robotic explorers. Although early visions of truly intelligent robots have yet to materialize (reasons for which will be outlined), there are nonetheless revolutionary developments which have attempted to address at least some of these earlier unperceived deficiencies. The current state of the art comprises a number of separate strands of research which provide components of robotic intelligence though no overarching approach has been forthcoming. The first question to be considered is the level of intelligent functionality required to support a long-duration starship mission. This will, at a minimum, need to be extensive imposed by the requirement for complex reconfigurability and repair. The second question concerns the tools that we have at our disposal to implement the required intelligent functions of the starship. These are based on two very different approaches - good old-fashioned artificial intelligence (GOFAI) based on logical theorem-proving and knowledge-encoding recently augmented by modal, temporal, circumscriptive and fuzzy logics to address the well-known "frame problem"; and the more recent soft computing approaches based on artificial neural networks, evolutionary algorithms and immunity models and their variants to implement learning. The former has some flight heritage through the Remote Agent architecture whilst the latter has yet to be deployed on any space mission. However, the notion of reconfigurable hardware of recent interest in the space community warrants the use of evolutionary algorithms and neural networks implemented on field programmable gate array technology, blurring the distinction between hardware and software. The primary question in space engineering has traditionally been one of predictability and controllability which online learning compromises. A further factor to be accounted for is the notion that intelligence is derived primarily from robot-environment interaction which stresses the sensory and actuation capabilities (exemplified by the behavioural or situated robotics paradigm). One major concern is whether the major deficiency of current methods in terms of lack of scalability can be overcome using a highly distributed approach rather than the hierarchical approach suggested by the NASREM architecture. It is contended here that a mixed solution will be required where a priori programming is augmented by a posteriori learning resembling the biological distinction between fixed genetically inherited and learned neurally implemented behaviour in animals. In particular, a biomimetic approach is preferred which exploits the neural processes and architecture of the human brain through the use of forward models which attempts to marry the conflicting requirements of learning with predictability. Some small-scale efforts in this direction will be outlined.

, , ,
JBIS - Journal of the British Interplanetary Society
Department of Mechanical and Aerospace Engineering

Ellery, A. (2009). Selective snapshot of state-of-the-art artificial intelligence & robotics with reference to the icarus starship. In JBIS - Journal of the British Interplanetary Society (Vol. 62, pp. 427–439).